ACADEMIA

Researchers from North Carolina State University have developed a new software tool to prevent performance disruptions in cloud supercomputing systems by automatically identifying and responding to potential anomalies before they can develop into problems.

Cloud computing enables users to create multiple "virtual machines" that operate independently, even though they are all operating on one large supercomputing platform. However, this approach can cause performance issues when a software bug, or other problem, in one virtual machine disrupts the entire cloud.

Now researchers have designed software that looks at the amount of memory being used, network traffic, CPU usage and other system-level data in a cloud supercomputing infrastructure to develop a definition of the wide range of behaviors that can be considered "normal." CPU usage is the amount of computing power being used at any given time. The program defines normal behavior for every virtual machine in the cloud, and can then look for deviations and predict anomalies that could affect the system's ability to provide service to users.

One advantage of this approach is that it does not require users to provide so-called "training data" about what constitutes abnormal behavior, which is important because training data are often difficult to obtain in production cloud systems. Moreover, this approach is also able to predict anomalies that have never been seen before.

If the program spots a virtual machine that is deviating from its normal behavior, it runs a "black box" diagnostic that can determine which metrics – such as CPU usage – may be affected, without exposing user data. This metric data can then be used to trigger the appropriate prevention system, which will address the deviation and prevent it from becoming a problem.

"If we can identify the initial deviation and launch an automatic response, we can not only prevent a major disturbance, but actually prevent the user from even experiencing any change in system performance," says Dr. Helen Gu, an assistant professor of computer science at NC State and co-author of a paper describing the research. "Also, it's important to note that this program does not access any user's individual information. We're looking only at system-level behavior."

The program is also lightweight, meaning it does not use much of the cloud's supercomputing power to operate. It is able to collect the initial data and define normal behavior much faster than existing approaches. Once it is up and running, it uses less than 1 percent of the CPU load and 16 megabytes of memory.

In benchmark testing, the program identified up to 98 percent of anomalies, which is much higher than the rate found in existing approaches. "It also had a 1.7 percent rate of false positives, meaning it triggered very few false alarms," Gu says. "And because the false alarms resulted in automatic responses, which are easily reversible, the cost of the false alarms is negligible."

Gu says her team's next step is to incorporate more detailed "white box" diagnostic tools into the software, so they can identify the software bugs causing any anomalies and correct them.